Anthropic Demonstrates Autonomous De Novo Protein Design and Chemical Analysis with Claude
Anthropic has published experimental results demonstrating Claude's ability to autonomously design de novo protein binders with physical wet-lab validation and automate complex analytical chemistry workflows. The findings show frontier LLMs acting as autonomous agents across computational biology and molecular characterization pipelines.
In the primary experiment, Anthropic evaluated Claude Mythos Preview and Opus 4.8 across a multi-arm de novo protein design campaign targeting 15 distinct protein structures. Independent wet-lab testing conducted by Adaptyv Bio and Twist Bioscience confirmed that Claude successfully generated functional binders for 14 out of the 15 targets, achieving hit rates significantly above traditional industry baselines.
Autonomous Binder Design and Wet-Lab Validation
De novo protein binder design involves generating synthetic amino acid sequences and 3D conformations capable of attaching tightly to a target macromolecule. The process traditionally requires weeks or months of iterative modeling and screening by specialized computational biologists.
Anthropic structured the campaign in Claude Science using two distinct configurations: a multi-target mode where Claude designed against all 15 targets in a single 48-hour session (allocating up to 12,500 NVIDIA H100 GPU-hours for folding models), and a single-target mode allocating 24 hours and up to 2,500 H100 GPU-hours per target. Given only high-level prompts and tool interfaces, Claude autonomously selected binding epitopes, orchestrated structure and sequence models, performed in-silico optimization cycles, and filtered candidates for solubility and novelty.
Across 1,320 total ordered designs, wet-lab validation yielded 354 confirmed binders. In multi-target mode, Claude Opus 4.8 and Mythos Preview achieved hit rates of 22.6 percent and 26.7 percent, respectively. When focused on a single target at a time, Mythos Preview achieved an overall hit rate of 35.1 percent, compared to the 10 to 15 percent hit rate typical in current computational campaigns. High-affinity binders were identified for at least six targets, with four targets yielding binders matching or exceeding the best affinities previously recorded in literature.
Claude failed on only one target, BBF-14, a de novo beta-barrel scaffold known in the research community for its lack of natural binding motifs.
Spectrometry Automation in Analytical Chemistry
In a companion study evaluating chemical characterization workflows, Anthropic tested Claude Opus 5 on interpreting raw nuclear magnetic resonance (NMR) and liquid chromatography-mass spectrometry (LC-MS) data. These datasets are used by synthetic chemists to verify molecular structure, identify degradation products, and measure sample purity.
Provided with raw data files and a brief prompt, Claude processed both analytical streams in parallel within 23 and 19 minutes, respectively. The model correctly determined hydrogen counts and measured compound purity at 96.4 percent, matching the contract lab's ground truth assessment of 96.33 percent. While human analytical reporting for such samples frequently requires multiple business days due to batch scheduling and manual peak integration, the model generated an auditable, fully interpreted report in under half an hour.
Anthropic indicated that while biological and chemical research tasks remain restricted in its frontier models, it plans to launch a dedicated access program to provide vetted life science researchers with specialized access to these capabilities.
